pt-wave-loop

Run the census script to list mirrored-but-incomplete units, pick 1-7 units capped near 15k golden lines prioritizing the largest gaps, then dispatch append-only translation agents. After integration, run the acceptance gate, commit, update the DEVLOG, and immediately start the next wave.

Use preprocessed output comparison: run g++ -E with the project's defines on both the HEAD version and the new file, strip line markers and carriage returns, and require every diff line to be blank. Then compile the changed file once with -O3 to rule out Release OOM, plus one Debug build as a regression check.

Static archive members are only extracted when they resolve an undefined symbol, so a file can compile into an archive yet never be linked. Check with nm --undefined-only and nm --defined-only to detect stubs satisfying the symbol first, wrong-archive registration, or static shadowing.

Only three conditions justify stopping: safety-critical behavior changes (motor brake, safety interlocks, real machine motion), reaching the form boundary where facade strategy is undecided, or compute quota exhaustion. Everything else, including test recalibration and agent retries, is pre-authorized to continue.

Run git log to see the last real commit, check git status for uncommitted work in progress, verify no background builds or workflows are still running, then read the RESUME marker at the end of DEVLOG.md. Never trust memory over git log, and never rerun completed waves.

Very large single-function files can cause 32-bit cc1plus to run out of memory at -O2 or -O3 while Debug compiles fine. The workaround is pinning that specific source file to -O1 via set_source_files_properties, and always measuring both build types after final integration.